DocVortex

A fast, multi-format document parsing and conversion engine.

DocVortex provides a complete, standalone document pipeline:

Document -> Unified Intermediate Representation -> Render / Export

Native inputs include text PDFs, DOC/DOCX, PPT/PPTX, XLS/XLSX, RTF, ODT/ODS/ODP, EPUB, HTML, OFD and CSV. Output formats include Markdown, HTML, LaTeX, DOCX, EPUB, PDF and structured content.

Native parsing runs without OCR or VLM inference services. PDF classification is an explicit document operation; native analysis does not silently classify the document or select another inference backend.

Install

pip install docvortex
docvortex convert report.pdf --format markdown --output output/report.md
docvortex classify report.pdf

Python 3.10–3.14 is supported. Native parsing does not require OCR/VLM inference services.

Parse once, export many times

import docvortex

result = docvortex.parse("report.pdf", keep_model_json=True)
result.export("output/report.md", output_format="markdown")
result.export("output/report.docx", output_format="docx")
result.export("output/report.epub", output_format="epub")
result.save_bundle("output/report.bundle")

# This works after the source document and its parsing process are gone.
restored = docvortex.load_bundle("output/report.bundle")
restored.export("output/report.pdf", output_format="pdf")

Bundles contain manifest.json, middle.json, optional model.json, and image assets. The loader verifies asset hashes. Missing external assets must be supplied before saving a portable bundle. Existing files are protected unless the caller explicitly sets overwrite=True.

Stage APIs

from docvortex.api import analyze, postprocess, render

analysis = analyze("report.pdf", page_range="1-5")
result = postprocess(analysis)
artifact = render(result.middle_json, "docx", assets=result.assets)
artifact.write("output/report.docx")

The stage API lives in docvortex.api. Root-level conveniences include parse, analyze, convert, postprocess_document, and render_artifact. The docvortex.render package also exposes the low-level renderers and their original string, bytes, dictionary, or list return values.

PDF page selections use 1-5, r1 and all; other native formats are parsed as whole documents. A caller-owned PDFDocument can be passed to analyze or parse and remains open afterward.

Explicit PDF classification

from docvortex.document.pdf import PDFDocument

with PDFDocument("report.pdf") as document:
    mode = document.classify()  # "txt" or "ocr"; no inference is started
    if mode == "txt":
        result = docvortex.parse(document)

Native analysis trusts the caller's choice and does not classify automatically. Applications can use the classification result to select their own OCR or inference service when a document requires it.

DocVortex JSON uses schema identity docvortex.model or docvortex.middle, schema version 2.0, and required metadata.file_suffix / metadata.producer. Definitions are in schemas/. Application-specific metadata belongs in extensions. See the shared JSON protocol and migration guide and the compatibility guide for existing application integrations and historical data formats. The HTML protocol describes DocVortex markers and semantic round trips.

Scope and development

PDF output is a semantic reflow of the document, not a lossless reproduction of the original page drawing instructions. Input support for PPTX/XLSX does not imply PPTX/XLSX output support. Rust implementation work is a future stage behind these public data and processing boundaries.

uv venv
uv pip install -e ".[test,dev]"
uv run --no-project python -m pytest -q
uv run --no-project ruff check src
uv run --no-project ruff format --check src
uv build

DocVortex project code is licensed under the MIT License.

See rendering ownership for the seven engine targets, MinerU Content List integration and public fragment helpers.